Tafel

An MCP server for Claude Code that turns "explain X with an animation" into a rendered explainer video on a chalkboard-style viewer page: the video plays muted on a loop, with a title, a summary and detailed, click-to-seek explanation blocks beneath it. Claude publishes the page as a claude.ai artifact, so you get a shareable link - no local file to manage.

Claude writes the animation code; your machine renders it.

Example

"explain how Fourier series build a square wave from sine waves, with an animation"

Final frame above - ▶ watch the full render to see it build up: one odd harmonic added at a time, the running sum snapping closer to a dashed target square wave with each term, the formula at the top growing in matching colors, ending on 25 terms and the Gibbs overshoot that never goes away. Rendered end-to-end from that one sentence - full interactive version (with the synced explanation blocks) here.

Related MCP server: Video MCP

Why it exists

Claude is a text model: on its own it can describe how a Fourier series or a TCP handshake works, or sketch a single static diagram - it cannot output a video, and generating even one genuinely coherent image showing motion or a process over time is out of reach. Tafel closes that specific gap. Claude writes a manim scene - code, which it's good at - and manim renders that code into real video frames locally. For anything where a moving picture explains more than words or a still image (a process unfolding, a wave building term by term, state changing over time), that's a capability Claude doesn't otherwise have.

Why it's cheap

AI video generators (Sora, Veo, Runway, Pika, ...) generate pixels directly: the model outputs the video itself, frame by frame, so cost and time scale with duration and resolution - a few seconds of video can cost real money and take minutes, every single time, even for small tweaks.

Tafel doesn't generate video at all. Claude only writes a manim Python script - a few hundred tokens of code describing shapes, text, and motion - plus the title/summary/ explanation text. Manim (a deterministic, open-source animation engine, the same one behind 3Blue1Brown's videos) then renders that script into actual pixels locally, using ffmpeg/cairo/pango on your own machine. The model never sees or produces a single frame.

Consequences:

• Cost is bounded by text, not video - a 30s clip and a 3-minute deep dive cost roughly the same in tokens; only your local render time differs.

• Iteration is nearly free - "slow down step 3" or "make the arrows curved" is a small code edit, not a full pixel regeneration from a prompt.

• Rendering is free and unlimited - it's your CPU/GPU, not a metered API, so there's no per-render fee and no external service dependency once the Python packages are installed.

• Fully offline after install - no network calls happen during a render; everything runs locally except the one Claude Code conversation.

Requirements

• macOS or Linux (Windows: install inside WSL2 and follow the Linux steps there)

• Claude Code

• Optional: LaTeX for mathematical formulas (several GB - see Troubleshooting)

Install

Unpack this folder anywhere permanent (e.g. ~/Documents/tafel), then:

cd tafel
./install.sh

The script installs the native dependencies (ffmpeg, cairo, pango) via Homebrew on macOS or your Linux package manager (apt/dnf/pacman), installs uv if missing, and registers the server with Claude Code as tafel (user scope, all projects). uv resolves the right Python version and the project's Python packages itself - no manual venv or Python install needed.

Manual equivalent, if you prefer:

# macOS
brew install ffmpeg cairo pango uv
# Linux (Debian/Ubuntu)
sudo apt-get install ffmpeg libcairo2-dev libpango1.0-dev pkg-config
curl -LsSf https://astral.sh/uv/install.sh | sh

chmod +x run-stdio.sh
claude mcp add tafel -s user -- "$PWD/run-stdio.sh"

Then open a new Claude Code session (servers load at session start) and check with /mcp that tafel shows as connected.

Usage

Just ask in any session:

explain how an SSL/TLS handshake works with an animation

Steer it with plain words - they map to the tool's settings below. Say things like "quick overview", "deep dive", "low quality, I'm iterating", or "give me a gif" and Claude picks the matching value; it defaults to standard depth and medium quality if you don't specify.

Settings

When the render finishes, Claude publishes the viewer page as a claude.ai artifact and shares the link - it does not open a local browser window. Iterate conversationally: "slow down step 3", "make the arrows curved", "redo this in high quality" - Claude edits the scene or settings and re-renders.

Environment variables

Set these before running install.sh, or edit them into run-stdio.sh after install (export TAFEL_TIMEOUT=600 before the exec uv run line), then reconnect the server with /mcp.

Output and cleanup

Each render briefly lands in the output directory (TAFEL_OUTPUT_DIR above) as <timestamp>_<Scene>.mp4 plus a matching self-contained .html viewer page (video embedded as a data URI). Claude deletes both right after publishing the artifact, so nothing accumulates on disk under normal use. If a session ends before cleanup runs, stale files may be left behind - safe to delete the directory's contents manually at any time.

Troubleshooting

• Tool missing in a session - servers load at session start; run /mcp to reconnect or open a new session. Same after updating server.py.

• Formula scenes fail ("latex: command not found" in the error) - install LaTeX: brew install --cask mactex-no-gui on macOS (~5 GB), or sudo apt-get install texlive texlive-latex-extra texlive-fonts-extra on Linux. Optional - only needed for Tex/MathTex formulas, not plain animations.

• Render timeout - see TAFEL_TIMEOUT under Environment variables above; raise it if high/fourk or deep_dive renders are cutting off.